Generative Models for Prediction of Non-B DNA Structures

Abstract

Motivation: Deep learning methods have been successfully applied to the tasks of predicting non-B DNA structures, however model performance depends on the availability of experimental data for training. Experimental technologies for non-B DNA structure detection are limited to the subsets that are active at the time of an experiment and cannot detect entire functional set of elements. Recently deep generative models demonstrated promising results in data augmentation approach improving classifier performance trained on augmented real and generated data. Here we aimed at testing performance of diffusion models in comparison to other generative models and explore the data augmentation approach for the task of non-B DNA structure prediction. Results: We tested denoising diffusion probabilistic and implicit models (DDPM and DDIM), Wasserstein generative adversarial network (WGAN) and vector quantised variational autoencoder (VQ-VAE) for the task of improving detection of Z-DNA, G-quadruplexes and H-DNA. We showed that data augmentation increased the quality of classifiers with diffusion models being the best for Z-DNA and H-DNA while WGAN worked better for G4s. Diffusion models are the best in diversity for all types of non-B DNA structures, WGAN produced the best novelty for G-quadruplexes and H- DNA. Since diffusion models require substantial resources, we showed that distillation technique can significantly enhance sampling in training diffusion models. When considering three criteria - quality of generated samples, sampling speed, and diversity, we conclude that trade-off is possible between generative diffusion model and other architectures such as WGAN and VQ-VAE. Availability: The code with conducted experiments is freely available at https://github.com/powidla/nonB-DNA- structures-generation.